Geospatial Domain

machine learning, maps, data visualisation

We built the map views in DataRobot’s AutoML platform, for exploring a dataset by location and for seeing where on the map a model’s predictions go wrong. They shipped as Location AI, a public beta in DataRobot’s 6.1 release in June 2020.

Location became a variable type of its own, next to numeric, categorical and date. A column could arrive as a Shapefile, GeoJSON, Esri File Geodatabase, Well-Known Text or PostGIS geometry, or as a latitude/longitude pair that DataRobot recognised and promoted to a location. Each one gets a Geospatial Map tab alongside its histogram.

DataRobot geospatial map exploring a dataset feature as a 3D hexagon-binned map of the Chicago area, colored and sized by average value

The default is a unique map: one mark per row, colored by value, with the exact coordinates on hover. Co-located rows are aggregated by min, max or average for a numeric feature, or by mode for a categorical one.

Unique map of the San Francisco Bay Area and Central Valley, each row drawn as a point colored by value, with the hovered point's coordinates shown under the legend

Past 20,000 unique geometries (or 50,000 vertices for lines and polygons) individual marks stop being legible, and the map switches to a kernel density view. Hexagon, grid and heat maps are also available. On the 3D maps, color and height encode two different things: the aggregated value, and how many rows fell into the bin.

3D grid map of the San Francisco Bay Area and Central Valley showing average values as blue bars by location

Hexagon map of the Central California coast from Bakersfield to Paso Robles, colored and sized by average value

Heatmap of the Los Angeles and San Diego region showing density of values across the map

Accuracy Over Space moves the same maps from the data to a trained model. It plots residuals, actuals or predictions for the validation, cross-validation or holdout partition, so a modeller can see where a regression model is systematically wrong, not just how wrong it is on average.

DataRobot map showing model residuals colored by polygon across the Utah region, for interpreting local prediction error

Accuracy Over Space map of Los Angeles and Orange County, points colored on a red-to-blue scale by average residual, with visualization, aggregation, data selection and metric type controls below

Announcement · Documentation